Summary
At an AI Engineer conference, Dustin Mihalik from Indeed shares production lessons from building MCP (Model Context Protocol) apps for Claude, ChatGPT, and their internal Career Scout job-search agent. Rather than covering MCP conceptually, the talk digs into the real engineering pitfalls his team hit when deploying these integrations at scale.
The central design rule Mihalik arrives at: anything rendered to the user must also be returned as structured data to the model. Without this, the AI treats the UI as a black box and cannot answer follow-up questions about displayed results — a mistake he has seen in multiple production MCP apps. He also covers how user interactions like button clicks and modal views need to be surfaced to the model via MCP’s updateModelContext method, and explains why getting Claude and ChatGPT to consistently generate clickable links required an unexpectedly large number of hours and evals to solve.
The talk concludes with a comparison of text-based MCP versus MCP apps for complex multi-search jobs, arguing that rich UI components give builders meaningful control over branding and layout that pure text responses cannot match. For teams considering porting an existing website or API into a chat interface like Claude or ChatGPT, this is one of the most concrete and experience-grounded guides available.
📺 Source: AI Engineer · Published September 09, 2026
🏷️ Format: Deep Dive
